SFMF-DETR: A Spatial-Frequency Collaborative Enhancement and Multiscale Feature Fusion Framework for Small Object Detection

초록

In remote sensing object detection tasks, accurately identifying tiny objects within vast and complex scenes is crucial. However, existing end-to-end frameworks face performance bottlenecks due to inadequate high-frequency detail preservation, fine-grained feature dilution during multiscale fusion, and spatial distortion from traditional upsampling. To address these issues, this article proposes SFMF-DETR, a spatial-frequency collaborative enhancement and multiscale feature fusion framework for small object detection. First, a spatial-frequency multiscale feature enhancement (SF-MFE) module is designed to augment high-frequency components of target boundaries via a dual-stream collaboration mechanism, alleviating spectral deficiency. Second, a cross-scale channel selection attention (CCSA) module implements fine-grained feature selection through adaptive weighting, preventing shallow features from being overwhelmed by deep semantics. Third, a soft-normalized upsampling alignment (SNUA) module preserves the structural integrity of tiny objects using energy normalization constraints. In addition, an Inner-SIoU loss function optimizes the regression branch through auxiliary scale strategies and direction-aware mechanisms, enhancing localization stability in dense scenarios. Experimental results on the VisDrone2019-DET, AI-TOD, and RSOD datasets demonstrate that SFMF-DETR achieves significant gains while maintaining real-time efficiency. Specifically, on VisDrone2019-DET, SFMF-DETR-R18 reaches an mAP(50) of 40.5% (+3.0%), while SFMF-DETR-R50 further reaches 42.1% (+3.4%). On AI-TOD, SFMF-DETR-R18 achieves 58.9% mAP50 (+2.8%) with a small-object average precision (AP) of 25.7% (+1.6%), and SFMF-DETR-R50 further improves these to 60.2% mAP(50) and 26.5% small-object AP. On RSOD, the model reaches 98.3% mAP50 (+3.3%), validating the effectiveness and generalization of the proposed framework across diverse remote sensing scenarios.

키워드